Automakers’ chips attract outside capital. Who can build an external business?

Key conclusions

Automaker-designed chips have reached the commercial test that follows production. Li Auto’s M100, NIO’s NX9031 and XPeng’s Turing are already installed in vehicles. Returns will depend on the number of models sharing each chip generation, its service life and continuing software improvements. For companies with proprietary algorithms and dependable deployment volumes, this can justify sustained investment. Joint development remains more economical for many other manufacturers.

Li Auto is still providing most of the money in its latest financing. Its October 16 announcement describes RMB1.625 billion in subscription agreements: RMB1 billion from Li Auto and RMB625 million from external investors. The business remains controlled and consolidated by Li Auto, with closing conditions outstanding. Outside capital supports funding and employee incentives; independent profitability will depend on orders.^[1]

Of these three manufacturers, XPeng has a clear external production reference. Volkswagen’s ID. UNYX 08 entered production in March with Turing chips and XPeng driving technology. NIO and Li Auto stand out for financing their chip subsidiaries. NIO’s September disclosure put its chip business’s latest post-money valuation at RMB12.25 billion. Investor backing and customer purchasing are progressing along different routes.^[2][3]

Deployment provides the clearest comparison

Sources: ^[3]–^[8]. Technical specifications are company disclosures. Development scope does not cover every third-party IP block, design tool, manufacturing or packaging activity. No common-workload public comparison is available across all three; the comparison therefore focuses on deployment and orders.

Li Auto and NIO retain control while funding chips separately

Li Auto is bringing investors into MHX Technology and introducing employee incentives. Assuming both incentive plans are fully used and the additional subscription right is fully exercised, Li Auto expects to hold 74.01%, investors 6.58% and incentive entities 19.41%. The additional right permits up to RMB500 million of investment before the end of 2026.^[1]

The structure addresses talent retention and investment duration. Engineers can participate in the value of the chip business while the parent retains direction over technology. This can help maintain teams across multiyear development cycles. Li Auto’s contribution represents approximately 60% of the signed subscription, indicating continuing parental support.

NIO moved earlier. Its February announcement disclosed RMB2.257 billion in financing agreements. September’s quarterly results then reported two agreements signed in June and August for a combined RMB493 million, implying a RMB12.25 billion post-money valuation. NIO’s subsidiary would hold 59.95% after completion, retaining control.^[2][9]

Sources: ^[1][2]. Amounts refer to signed agreements; ownership figures retain their stated completion conditions.

For investors, the automotive parent supplies an initial major customer but creates concentration risk. Internal purchasing supports iteration, while transfer pricing, vehicle sales and development priorities influence revenue. External customer revenue, delivery costs and cash collection should become operating measures as the subsidiary develops; otherwise it remains primarily a parent-funded development operation.

Control lets manufacturers plan models and silicon together and reduces exposure to changes in a supplier’s roadmap. Capital and management responsibilities remain within the group: chip failures, delays and weak deployment volumes still affect group returns.

The advantage comes from designing algorithms and silicon together

The three companies focus on AI computing through different designs. M100 emphasizes dataflow and reducing movement between memory levels. NIO combines neural and image processing with underlying software. XPeng designs specialized computing units around its models.^[5][6][7]

These decisions affect real engineering costs. In-car computing is constrained by power, cooling and response time. If memory transfer becomes the bottleneck, adding compute units may bring little improvement. Teams that can change model structure, compilers and data paths together can potentially perform the same task with less hardware, or run more capable models within the same power budget.

This fits manufacturers with a clear algorithm direction, strong software teams and reusable systems across models. They can determine supported operations, memory and interfaces early, then optimize software for deployed hardware throughout its life.

Silicon also fixes early assumptions for several years. A major change in algorithms can turn a specialized design into a constraint. Chip and model teams therefore need a shared roadmap, some general-purpose flexibility and development milestones ahead of critical vehicle decisions.

In a May interview, Li Auto CTO Xie Yan said M100 progressed from formal project approval in November 2022 to tape-out in 2024 and vehicle production in 2026—more than three years.^[10] Subsequent generations require work before the first generation has finished earning its return.

The reviewed material does not provide complete comparable chip-team headcounts or cumulative chip R&D for all three companies. Group research budgets cover vehicles, software and other activities; evaluating chip returns requires more detailed reporting.

Cost reduction depends on lifetime deployment

The economics combine upfront design, tools and verification with per-chip manufacturing, packaging and continuing support. Lifetime deployment spreads the fixed investment. Sharing a generation across more models lowers that burden per vehicle.

Flagship launches can establish a platform before broader deployment. If every model requires different silicon and separate software adaptation, even a large manufacturer may leave each individual program below economic scale.

The comparison should cover the entire domain controller. Fewer chips can reduce memory, power supply, cooling, wiring and validation costs. Conversely, larger models may require additional memory and cooling that consume part of the purchasing savings.

Over the next one to two years, the strongest returns should accrue to manufacturers that deploy one computing platform across several mainstream models while maintaining common software. A chip serving only a low-volume flagship faces much tougher amortization. Companies without their own algorithms or sufficient volume can better protect launch timing and budgets by purchasing mature platforms with access to interfaces, tools and joint optimization.

XPeng’s external business starts with a deliverable system

ID. UNYX 08 combines Turing chips with XPeng driving technology. The customer is buying a technology package that can reach vehicle production. XPeng reports 24 months between the joint development agreement and production.^[3]

This reduces the customer’s integration work across hardware, software and vehicle validation, while creating opportunities beyond chip supply. Public disclosures do not separate chip prices, licensing and engineering fees for this project, so Turing’s external margin cannot be isolated.

Selling to more manufacturers also requires support for different sensors and vehicle platforms, customer-specific software schedules, usable development tools, fault responsibility and long-term supply. Customers need assurance that their algorithms and data remain separate from the supplier’s automotive parent.

Separate financing gives Li Auto and NIO room to build those organizations. The next commercial evidence will be additional production customers and recurring purchases. Robotics can expand applications, but more fragmented requirements and project volumes suggest automotive deployment will remain the main source of near-term scale.

Suppliers face different kinds of pressure

NVIDIA initially faces internal substitution in some premium vehicle-computing programs. The three automakers now have deployed alternatives. NVIDIA nevertheless announced in March that BYD, Geely, Isuzu and Nissan would use DRIVE Hyperion for Level 4 development.^[11] Tools, verification and customer coverage become more important competitive strengths as pricing based primarily on compute specifications becomes harder to sustain.

Qualcomm’s cockpit position faces less immediate disruption. The new L9 retains Qualcomm alongside M100, and ET9 uses 8295P.^[4][8] Displays, audio, applications and compatibility create a separate engineering and commercial decision for cockpit replacement.

Horizon faces stronger pricing competition but also opportunities in joint development. Its July disclosure reported Journey 6 design wins with more than 25 automakers and over 100 models, with more than 40 variants in production. Volkswagen’s joint development through Carizon illustrates demand for this route.^[12] Broad customer coverage spreads tooling and research costs; automaker-owned suppliers offer tighter integration with their own models.

Over the next two to three years, the cost of a computing system performing a given driving task is likely to fall through specialization, competition and integration. Flagship chip prices may still reflect larger models, memory and safety requirements. Public procurement data does not support a uniform percentage price forecast.

Vehicle manufacturers should increasingly compare model performance, power use, total controller cost and post-launch support. Chip companies will be judged by how reliably they can serve more models and customers.

Proprietary SoCs will first strengthen a small number of leading automakers’ products, before supporting broad external sales. XPeng already has an external-brand production reference. Li Auto and NIO need to convert separate financing into customers and revenue. For most manufacturers, joint development offers a practical way to differentiate while concentrating capital on the products and software where they can win.